[IMPALA-5907] Increased codegen time on specific queries in Impala 2.6

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Information Provided
• Affects Version/s: Impala 2.6.0
• Component/s: Backend
• Target Version: Impala 2.11.0
• Epic Color: ghx-label-9
• Assignee: Bikramjeet Vig
• Reporter: Balazs Jeszenszky
• Created: 07/Sep/17 15:13
• Updated: 26/Oct/17 22:22
• Resolved: 26/Oct/17 22:22

Description

This also repro's in latest version, but it was introduced in CDH5.8 / Impala 2.6. It happens with a large amount of OR predicates over a wide table. I've attached two profiles, the codegen part is interesting:

2.5:

CodeGen:(Total: 28s173ms, non-child: 28s173ms, % non-child: 100.00%)
           - CodegenTime: 507.007ms
           - CompileTime: 16s150ms
           - LoadTime: 0.000ns
           - ModuleBitcodeSize: 1.84 MB (1925128)
           - OptimizationTime: 11s791ms
           - PrepareTime: 202.002ms

2.6:

CodeGen:(Total: 11m36s, non-child: 11m36s, % non-child: 100.00%)
           - CodegenTime: 40.000ms
           - CompileTime: 15s580ms
           - LoadTime: 0.000ns
           - ModuleBitcodeSize: 2.19 MB (2300548)
           - NumFunctions: 3.51K (3514)
           - NumInstructions: 74.28K (74282)
           - OptimizationTime: 11m20s
           - PrepareTime: 169.000ms

Issue Links

• relates to: IMPALA-5980 Upgrade to LLVM 5.0.1